An intelligent diagnosis method for gearbox faults based on motor stator current signals

Abstract

The present invention discloses an intelligent diagnosis method for gearbox faults based on motor stator current signals, comprising the following steps: 1, constructing a CNN model sample library; 2, constructing a comb filter; 3, building a CNN model, wherein the input is the motor stator current signal corresponding to the gearbox; the output is the gearbox fault type; each convolution kernel of the first convolution layer is a comb filter; 4, training the CNN model; and 5, diagnosing the gearbox fault. The present invention utilizes the motor stator current signal to identify the health status of the gearbox, without the need to install external sensor equipment on the gearbox, thereby reducing the dependence of the diagnostic results on the sensor layout and reducing the impact of environmental noise on signal acquisition. In addition, the present invention does not require pre-processing operations such as denoising and removing the fundamental frequency component of the current signal on the collected motor stator current signal, and can perform end-to-end fault diagnosis tasks.

Description

Technical Field

[0001] The present invention relates to the field of mechanical equipment fault diagnosis, and in particular to an intelligent gearbox fault diagnosis method based on motor stator current signals. Background Art

[0002] With the rapid development of modern industry, rotating machinery is gradually becoming larger, more complex, and more automated. Because rotating machinery often operates in harsh working environments at high speeds and under heavy loads, it is prone to equipment failure. Equipment failure can cause serious production accidents, potentially leading to economic losses for the manufacturer or even life-threatening situations for production personnel. Statistics show that gearbox failures account for approximately 80% of all mechanical transmission equipment failures, making gearbox fault diagnosis crucial. Existing fault diagnosis technologies are largely based on vibration signal analysis, but vibration sensors are expensive and difficult to install. Furthermore, the sensor's installation location and ambient noise significantly impact the acquisition of vibration signals, which in turn affects the final diagnostic results.

[0003] In recent years, current signal analysis methods that do not require additional sensors have gained attention. Because motors inherently possess sensory properties, changes in their stator current can reflect the dynamic characteristics of the machinery they drive. When a fault occurs in the motor or a transmission component, it causes torque fluctuations. This fluctuating torque triggers changes in the rotating electromagnetic field in the motor's air gap, modulating the fault-related frequency characteristics into the stator current signal. Therefore, current signal analysis can be used to diagnose faults in rotating machinery.

[0004] While current signal analysis overcomes the disadvantage of vibration signal analysis, which requires additional sensors, when diagnosing faults in transmission components other than the motor, the long transmission path results in a relatively weak fault component in the motor's stator current. Furthermore, the presence of the current fundamental frequency, slot harmonics, and inherent eccentricity harmonics acts as noise frequency components that overwhelm the fault information, resulting in a low signal-to-noise ratio (SNR) in the current signal. Accurate fault classification based on this low SNR presents a challenge for fault diagnosis. Many researchers have combined signal processing techniques with machine learning algorithms, using them to suppress noise components unrelated to the fault and extract time-domain and frequency-domain features from the signal. These features are then fed into a machine learning algorithm as a classifier for feature classification, completing the entire fault diagnosis process. However, these methods rely on expert experience to manually screen fault features, resulting in poor generalization performance of the fault diagnosis model and low diagnostic efficiency.

[0005] Intelligent fault diagnosis methods based on deep learning algorithms can adaptively extract fault features from raw signals without manual intervention, and integrate the feature extraction and classification processes to achieve an end-to-end process from data to diagnostic results. Convolutional Neural Networks (CNN), as an important development branch of deep learning theory, have been applied by many scholars in the field of fault diagnosis. CNN utilizes its stacked convolutional layers, activation layers, pooling layers, and other structures to layer-by-layer explore the complex nonlinear relationship between mechanical signals and faults, achieving a high fault diagnosis accuracy. However, current intelligent fault diagnosis methods based on current signal analysis often require preprocessing of the current signal, removing the power supply frequency component of the current signal before inputting it into the CNN model for fault identification, which makes it impossible to achieve true end-to-end fault diagnosis. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide an intelligent gearbox fault diagnosis method based on the motor stator current signal. The intelligent gearbox fault diagnosis method based on the motor stator current signal can accurately diagnose the gearbox fault using the original data of the motor stator current signal without the need for signal preprocessing technology.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A gearbox fault intelligent diagnosis method based on motor stator current signals includes the following steps.

[0009] Step 1: Build a CNN model sample library, which includes the following steps.

[0010] Step 1-1. Construct gearbox fault types: The gearbox of a rotating machine includes input gears and output gears that mesh with each other. Design a number of input gears with different fault types and b number of output gears with different fault types. By cross-combining a number of input gears and b number of output gears, a*b number of gearbox faults with different fault types are obtained. Where a and b are both natural numbers.

[0011] Step 1-2, collecting the motor stator current: The gearbox for each fault type constructed in step 1-1 is sequentially installed in a rotating machine containing a motor; the rotating machine is operated, and a current sensor is used to collect the stator current of the motor at a set frequency; and the motor stator current signal of the gearbox for each fault type is obtained.

[0012] Step 1-3, construct a CNN model sample library: the stator current signal of the gearbox motor for each fault type obtained in step 1-2 constitutes the input sample library of the CNN model; the a*b types of gearbox fault types obtained in step 1-1 form the output sample library of the CNN model; the input sample library and the output sample library together form the CNN model sample library.

[0013] Step 2: Construct a comb filter, which specifically includes the following steps:

[0014] Step 2-1: Characterize the fault characteristic frequency f in the current signal fallt , the specific expression is:

[0015] f fault =|f s ±kf fp |

[0016] When the gearbox is a single-stage gearbox, the following applies:

[0017] f fp =f r，in or f fp =f r，olt

[0018] Where, f s The frequency of power supply to the motor; f fp The rotation frequency of the faulty gear.

[0019] k is the harmonic order, a natural number.

[0020] f r，in is the input gear rotation frequency.

[0021] f r，olt is the output gear rotation frequency.

[0022] Step 2-2, characterize the noise frequency: the noise frequency includes f s and f s ±c; where c is the torque fluctuation frequency caused by non-gearbox faults.

[0023] Step 2-3, construct a comb filter: The passband of the comb filter can make the fault characteristic frequency f characterized in step 2-1 fallt By doing so, the stopband of the comb filter can suppress the noise frequency characterized in step 2-2.

[0024] Step 3. Build a CNN model: The CNN model includes an input layer, a hidden layer, and an output layer. The input layer is the motor stator current signal corresponding to the gearbox. The output layer is the gearbox fault type. The hidden layer includes several convolutional layers. The first convolutional layer has several convolution kernels. Each convolution kernel of the first convolutional layer is the comb filter constructed in step 2.

[0025] Step 4: Train the CNN model: Use the CNN model sample library constructed in step 1 to train the CNN model sample library built in step 3 to obtain a CNN model with fault diagnosis capabilities.

[0026] Step 5: Gearbox fault diagnosis, specifically including the following steps:

[0027] Step 5-1, obtaining the actual motor stator current signal: using a current sensor to collect the motor stator current corresponding to the gearbox to be tested at a set frequency; obtaining the actual motor stator current signal of the gearbox to be tested.

[0028] Step 5-2, gearbox fault diagnosis: The actual motor stator current signal obtained in step 5-1 is used as input and substituted into the CNN model with fault diagnosis capability obtained in step 4. The output is the fault type of the gearbox to be detected.

[0029] In step 1-1, a=b=4, and the four different fault types of the input gear are severe pitting, moderate pitting, mild pitting, and no fault; among them, severe pitting, moderate pitting, and mild pitting mean that there are pitting on the input gear, and the number of pitting decreases gradually.

[0030] The four different fault types of the output gear are severe wear, moderate wear, mild wear and no fault; among them, severe wear, moderate wear and mild wear mean that at least one tooth of the output gear is worn, and the wear thickness gradually decreases.

[0031] In step 2-1, k is set according to the current amplitude observed in the motor stator current signal corresponding to the gearbox in step 1.

[0032] In step 2-3, the comb filter is constructed using the Fourier kernel function; the time domain function of the comb filter is about time t, f h，m and f l，m function; f h，m is the high cutoff frequency of the mth comb tooth, f l，m is the lower cutoff frequency of the mth comb tooth, 1≤m≤M; M is the number of comb teeth of the comb filter; the stopband bandwidth between adjacent comb teeth is marked as f d .

[0033] f h，m 、f i，m and f d All of them are self-learning parameters. By adjusting the high cutoff frequency f of the first comb tooth h，1 , the low cutoff frequency f of the first comb tooth i，1 and f d Initial values are set for both, and self-learning is performed during the CNN model training in step 4.

[0034] f i，1 The initial value range is [0,80%×f s / 2];f d The initial value f d0 The frequency resolution of not less than 2 comb filters; f h，1 The initial value range is [f d0 ,80%×f s / 2 +f d0 ].

[0035] In step 3, the number of convolution kernels in the first convolution layer does not exceed 45, and the average accuracy of gearbox fault diagnosis in step 5 is not less than 99%.

[0036] The number of convolution kernels in the first convolution layer is 40.

[0037] Before the gearbox fault diagnosis in step 5, first set the motor power supply frequency f corresponding to the gearbox to be tested. s The motor power supply frequency f in step 2 s Compare them. If the two are the same, proceed to step 5. Otherwise, according to the motor power supply frequency f corresponding to the gearbox to be tested, s , repeat steps 1 to 4 to train a new CNN model; then proceed to step 5 based on the new CNN model.

[0038] In step 2-3, the comb filter is truncated using a Hamming window.

[0039] The present invention has the following beneficial effects:

[0040] 1. The health status of the gearbox is identified using the motor stator current signal. There is no need to install external sensor equipment on the gearbox, which reduces the dependence of the diagnostic results on the sensor layout and the impact of environmental noise on signal acquisition.

[0041] 2. The present invention does not require pre-processing operations such as noise reduction and removal of the fundamental frequency component of the collected motor stator current signal, and can perform end-to-end fault diagnosis tasks.

[0042] 3. Each comb filter in the comb filter layer of the improved CNN created by this invention can mine fault information in different frequency bands of the motor stator current signal. The comb filter waveforms can be adaptively adjusted during training to adjust the passband and stopband ranges of each comb filter. This enhances the fault feature extraction capability of the entire improved CNN and helps the model achieve higher fault identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1This is a flow chart of a gearbox fault intelligent diagnosis method based on motor stator current signals of the present invention.

[0044] Figure 2 This is a technical diagram of the leave-out method.

[0045] Figure 3 This is a technical diagram of the random fixed-length segmentation method.

[0046] Figure 4 It is a structural diagram of the improved CNN model in the present invention.

[0047] Figure 5 2 is a model training accuracy curve diagram of an embodiment of the present invention.

[0048] Figure 6 2 is a graph showing the model training loss value according to an embodiment of the present invention.

[0049] Figure 7 4 is a confusion matrix diagram of the fault diagnosis classification results of an embodiment of the present invention.

[0050] Figure 8 This is the spectrum diagram of the measured motor stator current signal and its low-frequency band when the gearbox is faulty.

[0051] Figure 9 This is a t-SNE dimension reduction visualization diagram of the fault diagnosis classification result of an embodiment of the present invention.

[0052] Figure 10 It is a comparison diagram of the original spectrum and filtered spectrum of the simulation signal; among them, (a) is the original spectrum diagram of the simulation signal; (b) is the filtered spectrum diagram of the simulation signal after being filtered by a comb filter.

[0053] Figure 11 Figure 1 is a schematic diagram of the spectrum waveforms of the four comb filters and their corresponding feature maps before and after model training. Figures a and b are the spectrum waveforms of the first comb filter and its corresponding feature map before and after model training, respectively; Figures c and d are the spectrum waveforms of the second comb filter and its corresponding feature map before and after model training; Figures e and f are the spectrum waveforms of the third comb filter and its corresponding feature map before and after model training; and Figures g and h are the spectrum waveforms of the fourth comb filter and its corresponding feature map before and after model training. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and specific preferred embodiments.

[0055] In the description of the present invention, it should be understood that the terms "left side," "right side," "upper," "lower," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Terms such as "first" and "second" do not indicate the importance of components and therefore should not be construed as limitations on the present invention. The specific dimensions used in this embodiment are intended only to illustrate the technical solution and do not limit the scope of protection of the present invention.

[0056] like Figure 1 As shown, a gearbox fault intelligent diagnosis method based on motor stator current signal includes the following steps.

[0057] Step 1: Build a CNN model sample library, which includes the following steps.

[0058] Step 1-1: Constructing gearbox fault types

[0059] The above-mentioned rotating machinery includes a three-phase asynchronous motor, a first-stage gearbox and a load loading device connected in sequence. A Hall current sensor is installed on phase A of the three-phase asynchronous motor for collecting the motor stator current signal when the gearbox is in different health states.

[0060] The above-mentioned first-stage gearbox includes an input gear and an output gear that mesh with each other; a types of input gears with different fault types and b types of output gears with different fault types are designed; wherein a and b are both natural numbers. In this embodiment, preferably a=b=4, as shown in Table 1 below.

[0061] Table 1 Gearbox input / output gear details

[0062]

[0063] As shown in Table 1, the four different fault types of the input gear are severe pitting, moderate pitting, mild pitting, and no fault; among them, severe pitting, moderate pitting, and mild pitting mean that the input gear preferably has pitting with a diameter of 0.5 mm, and the number of pitting decreases gradually, preferably 96, 42, and 15, respectively.

[0064] As shown in Table 1, the four different fault types of the output gear are severe wear, moderate wear, mild wear and no fault; among them, severe wear, moderate wear and mild wear mean that at least one tooth of the output gear is worn, and the wear thickness gradually decreases, preferably in the order of: 0.15mm, 0.1mm and 0.05mm.

[0065] By cross-combining a type of input gear and b type of output gear, a*b types (preferably 16 types) of gearbox faults with different fault types are obtained;

[0066] Step 1-2, collecting the motor stator current: install the gearbox for each fault type constructed in step 1-1 in the rotating machinery containing the motor in turn; the rotating machinery is working, and a current sensor (Hall current sensor) is used to collect the stator current of the motor at a set frequency; obtain the motor stator current signal of the gearbox for each fault type.

[0067] In a specific embodiment, when the motor power supply frequency is 30 Hz and the motor stator current signal is collected, the sampling frequency of the signal is preferably 6400 Hz, and the sampling time is preferably 20 s.

[0068] Steps 1-3: Build a CNN model sample library.

[0069] The stator current signal of the gearbox motor for each fault type obtained in step 1-2 constitutes the input sample library of the CNN model; the a*b types of gearbox fault types obtained in step 1-1 form the output sample library of the CNN model; the input sample library and the output sample library together form the sample library of the CNN model.

[0070] In a specific embodiment, Figure 2 As shown, the holdout method is preferably used to divide the CNN model sample library into training and test sets in a ratio of 4:1. The total length of the current signal sampling data is 128,000, of which the first 80% of the length (102,400) is used to prepare the training set samples, and the last 20% of the length (25,600) is used to prepare the test set samples.

[0071] In a specific embodiment, Figure 3 As shown, the random fixed-length segmentation method is preferably used to obtain samples. For example, the training set samples are prepared with a current signal data length of 102400. Considering that the sample length must cover at least one gearbox rotation cycle, and since the output gear of the gearbox rotates at 10.7 Hz, every 0.1 second of data covers a complete rotation cycle, each sample length is set to 3200 sampling points, corresponding to 0.5 seconds of data covering five complete rotation cycles. The sampling window is 3200. A random index is selected within the interval [1, 102400-3200] to extract the sample, resulting in a dimension of 1×3200 for each sample. This enhances the randomness of the data and thus increases the robustness of the model. For each gearbox fault type, 400 samples are selected as the training set, and 100 samples are selected as the test set, resulting in a total of 6400 samples in the training set and 1600 samples in the test set.

[0072] Step 2: Construct a comb filter, which specifically includes the following steps:

[0073] Step 2-1: Characterize the fault characteristic frequency f in the current signal fallt , the specific expression is:

[0074] f fault =|f s ±kf fp |

[0075] When the gearbox is a single-stage gearbox, the following applies:

[0076] f fp =f r，in or f fp =f r，olt

[0077] Where, f s The frequency of power supply to the motor; f fp The rotation frequency of the faulty gear.

[0078] f r，in is the input gear rotation frequency.

[0079] f r，out is the output gear rotation frequency.

[0080] The above fault characteristic frequency has a comb spectrum characteristic with multiple sidebands centered on the power supply frequency and interval frequencies equal to the fault gear rotation frequency; and research has shown that the fault characteristic frequency will appear in the same form on both sides of the integer multiple harmonics of the power supply frequency. In order to more intuitively display the distribution of the fault characteristic frequency, we collected the measured motor stator current signal when the input gear and output gear of the gearbox were faulty and plotted its low-frequency spectrum diagram, as shown in the figure below. Figure 8 shown.

[0081] In this embodiment, the input drive gear of the gearbox is preferably a 55-tooth spur gear, and the output gear is preferably a 75-tooth spur gear. The power supply frequency of the motor is f s =30Hz, the rotation frequencies of the input gear and output gear are f r，in =14.55Hz and f r，olt=10.7Hz. It can be found that the rotation frequency of the two faulty gears and their harmonics appear as modulation frequencies on both sides of the motor power supply frequency, and the fault characteristic frequency formula is verified. In this measured signal, the theoretical fault characteristic frequency generated by the input gear rotation frequency and its harmonics can be observed until k=3, but the theoretical fault characteristic frequency generated by the output gear rotation frequency can only be observed when k=1. The theoretical fault characteristic frequency generated by higher-order harmonics cannot be observed due to their small amplitude. In addition, compared with the amplitude of the power supply frequency, the amplitude of the fault characteristic frequency is very small, and the spectrum diagram needs to be partially magnified to observe the fault characteristic frequency. This is exactly the manifestation of the weak fault characteristics in the motor stator current signal.

[0082] Therefore, suppressing the motor power supply frequency and other noise frequencies while retaining the fault characteristic frequency can effectively improve the signal-to-noise ratio of the motor stator current signal, thereby increasing the accuracy of the motor stator current-based fault diagnosis method. As a multi-band bandpass filter, the comb filter consists of many passbands and stopbands with equal frequency intervals. It only allows frequency components in a specific frequency band to pass and suppresses the interference of noise components. Therefore, the comb filter meets our needs of suppressing the motor power supply frequency and other noise frequencies while retaining the fault characteristic frequency. It can be used as a predefined FIR filter to replace the convolution kernel of the first layer of the standard CNN to extract the fault characteristic frequency in the motor stator current signal and suppress the noise frequency, thereby improving the fault diagnosis accuracy of the CNN.

[0083] Step 2-2, characterize the noise frequency: the noise frequency includes f s and f s ±c; where c is the torque fluctuation frequency caused by non-gearbox faults, such as assembly errors.

[0084] Step 2-3, construct a comb filter: The passband of the comb filter can make the fault characteristic frequency f in the current signal characterized in step 2-1 fault By doing so, the stopband of the comb filter can suppress the noise frequency characterized in step 2-2.

[0085] There are many ways to construct a comb filter, such as Morlet wavelet, harmonic wavelet, and Fourier kernel function. Since the Fourier kernel function is a flat-top low-pass filter with an amplitude of 1 in the frequency domain, based on the superposition of Fourier transform, a comb filter can be easily constructed using multiple Fourier kernel functions. Therefore, this patent uses the Fourier kernel function to construct a comb filter. First, the frequency domain expression of the ideal comb filter is given as:

[0086]

[0087] Where M is the number of teeth of the comb filter, fh，m is the high cutoff frequency of the mth comb tooth, f i，m is the lower cutoff frequency of the mth comb tooth, and the stopband bandwidth f between adjacent comb teeth d =f i，m+1 -f h，m , and rect(·) is a rectangular function defined as:

[0088]

[0089] Therefore, the time domain expression of the comb filter is:

[0090]

[0091] In order to idealize the spectral shape of the filter using a finite number of parameters, a Hamming window is applied to h(t):

[0092] h W (t,f h，m , f i，m )=h(t,f h，m , f i，m )·w(t)

[0093] In the present invention, the window function w(t) is a Hamming window, and its mathematical expression is:

[0094] w(t)=0.54-0.46·cos(2πt / L)

[0095] Where L is the length of the comb filter.

[0096] To intuitively demonstrate the effectiveness of the comb filter, this paper uses simulation signals to verify the comb filter's ability to extract gear fault features from the stator current signal. Based on the formula for the fault characteristic frequency, we establish a gearbox fault simulation current signal sim(t):

[0097]

[0098] Among them, 60Hz is the simulated stator current supply frequency, 28Hz and 8Hz are the simulated fault gear rotation frequencies, is a randomly set interference signal, which can be generated by the inherent torque fluctuation caused by assembly error. At this time, 32Hz, 52Hz, 68Hz, and 88Hz in this simulation signal are fault characteristic frequencies, and 60Hz, 39Hz, and 81Hz are noise frequencies. We use h W(t) Construct a comb filter. Considering the distribution of noise frequency and fault characteristic frequency in the spectrum of this simulation signal, the number of comb teeth in this simulation experiment is set to 4. In order to make the passband of the comb filter pass the fault characteristic frequency and the stopband suppress the noise frequency, after calculation, the low cutoff frequency of the first comb tooth is set to 26Hz, and the passband and stopband width of the comb filter are both set to 10Hz. Use this comb filter to perform convolution operation on the simulation signal in the time domain, and draw the original spectrum of the simulation signal and the spectrum of the simulation signal after filtering as shown below. Figure 10 shown.

[0099] from Figure 10 It can be found that after the simulation signal is filtered, the simulated power supply frequency of 60Hz and the noise frequencies of 39Hz and 81Hz are significantly suppressed, while the fault characteristic frequencies of 32Hz, 52Hz, 68Hz, and 88Hz are well preserved. This result shows that the signal filtered by the comb filter has a higher signal-to-noise ratio than the original signal, which helps to improve the accuracy of fault diagnosis.

[0100] Step 3: Build a CNN model.

[0101] like Figure 4 As shown in the figure, the constructed CNN model includes an input layer, a hidden layer and an output layer; the input layer is the stator current signal of the motor corresponding to the gearbox; the output layer is the gearbox fault type; the hidden layer includes several convolution layers; among them, the first convolution layer has several convolution kernels; each convolution kernel of the first convolution layer is a comb filter constructed in step 2, and this improved convolution layer is called a comb filter layer.

[0102] When the first convolution kernel of the standard CNN is replaced by a comb filter, a practical problem is taken into consideration: since the power supply frequency of the motor will not be set to the same value every time, if the passband and stopband frequency bands of the comb filter are fixed, then the comb filter cannot successfully perform the function of suppressing noise frequency and extracting fault features every time. In order to enable the stopband of the comb filter to smoothly suppress the power supply frequency and other noise frequencies of the motor, and to make the improved CNN have higher generalization performance, we need to enable the comb filter to be continuously trained and learned during the training process of the CNN model to adapt to the current signals under different power supply frequencies. Fortunately, h W is differentiable, so we can set the parameter f i，1 , f h，1 , f d The learnable parameters set as comb filters are updated and optimized along with other parameters in the model during model training. This update and optimization of the learnable parameters in the comb filters allows them to adaptively adjust the comb filter waveforms during model training to adjust the passband and stopband ranges of each comb filter.

[0103] Furthermore, considering that the trainable parameters in the neural network need to be initialized, the parameter f i，1 In theory, it is possible

[0104] [0,f s / 2] range initialization, f s Indicates the sampling frequency of the input signal. Considering that the comb filter will occupy a certain bandwidth in the frequency domain, it is selected in [0,80%×f s / 2] to initialize f within the scope i，1 In this specific embodiment, this range is [0, 2560]. Since the sampling frequency is 6400Hz, the length of the comb filter is 1025, and the frequency resolution can only reach 6.24Hz, in order to prevent the spectral peaks of each comb filter from overlapping to form a comb spectrum, the frequency band interval between the high and low cutoff frequencies of each comb tooth needs to reach 2 frequency resolutions or more, and the passband width and stopband width need to be greater than 12.48Hz. Therefore, it is necessary to make the passband and stopband width of the comb filter meet this requirement during initialization. Therefore, this patent is in [20, 80% × f s Initialize f within the range of [ / 2 + 20] h，1 , so that the passband bandwidth is initialized to 20. In this specific embodiment, the specific value of this range is [20, 2580], and the stopband width parameter f d Initialization value f d0 Also set to 20.

[0105] In addition, the number of convolution kernels in the first convolution layer does not exceed 45. Conventional CNNs typically use 64 convolution kernels in the first convolution layer. This is because, to a certain extent, a greater number of convolution kernels means more different types of features are extracted and the CNN's feature extraction capability is stronger, but this also reduces the computational efficiency of the CNN. The present invention balances the feature extraction capability and computational efficiency of the CNN, optimizing the number of convolution kernels in the first convolution layer to 40. While significantly reducing the number of convolution kernels in the first convolution layer, it also ensures an average gearbox fault diagnosis accuracy of no less than 99%.

[0106] Furthermore, the hidden layer also includes a Batch Normalization layer (BN layer), an activation layer, a pooling layer, a global average pooling layer, a fully connected layer, and a Softmax classifier. In a specific embodiment, the specific architecture diagram of the improved CNN model constructed in this way is shown in Table 2 below.

[0107] Table 2 Improved CNN model architecture

[0108]

[0109]

[0110] Step 4: Train the CNN model: Use the CNN model sample library constructed in step 1 to train the CNN model sample library built in step 3 to obtain a CNN model with fault diagnosis capabilities.

[0111] The parameters for model training are preferably set as follows: batch size = 128 samples used for model training, maximum number of model training epochs = 100, initial learning rate α0 set to 0.05, learning rate decay rate γ set to 0.9, and learning rate decay once every 20 training cycles. When the model training reaches the 100th cycle, the training is stopped. The model with the highest diagnostic accuracy on the test set samples in 100 rounds of training is selected as the final diagnostic model. The accuracy curve generated during the model training process is shown in the figure below. Figure 5 As shown, the loss value curve is as follows Figure 6 shown.

[0112] Figure 11 The spectral waveforms of the four comb filters before and after model training and the spectral waveforms of their corresponding feature maps are shown. It can be observed that the comb filter can learn from the current signal the f that is conducive to extracting the fault characteristic frequency and suppressing the noise frequency. l，1 , f h，1 , f d value, thereby adaptively adjusting the range of the passband and stopband: the spectrum of the first filter shifts to the left after training, and the spectrum of its feature graph reflects that the 30 Hz motor power supply frequency is suppressed, and the adjacent fault characteristic frequency is retained; the passband and stopband of the second filter are widened after training, the spectrum of the third filter shifts to the right after training, and the passband and stopband of the fourth filter are narrowed after training. The spectrum of their feature graphs also changes with the changes of the comb filter.

[0113] Step 5: Gearbox fault diagnosis

[0114] Before diagnosing the gearbox fault, first set the motor power supply frequency f s The motor power supply frequency f in step 2 s Compare them. If the two are the same, proceed to step 5. Otherwise, according to the motor power supply frequency f corresponding to the gearbox to be tested, s , repeat steps 1 to 4 to train a new CNN model; then proceed to step 5 based on the new CNN model.

[0115] Step 5-1, obtaining the actual motor stator current signal: using a current sensor to collect the motor stator current corresponding to the gearbox to be tested at a set frequency; obtaining the actual motor stator current signal of the gearbox to be tested.

[0116] Step 5-2, gearbox fault diagnosis: The actual motor stator current signal obtained in step 5-1 is used as input and substituted into the CNN model with fault diagnosis capability obtained in step 4. The output is the fault type of the gearbox to be detected.

[0117] In a specific embodiment, the accuracy is preferably used as an indicator for evaluating the performance of the model, which is defined as:

[0118]

[0119] Among them, Acc is the accuracy rate, TP indicates accurate classification of various types of gear faults, TN indicates accurate classification of healthy gears, FP indicates that healthy gears are misclassified into various types of faults, and FN indicates that gear faults are misclassified into other fault types or healthy gears.

[0120] The constructed test set samples were used to verify the fault diagnosis capability of the improved CNN proposed in this paper. To demonstrate the superiority of the improved CNN, a traditional CNN with the same network architecture as the improved CNN was used as a control algorithm for comparative analysis. To reduce the influence of accidental factors, five different test set samples were constructed using the random fixed-length segmentation method. The two algorithms were tested five times. The test results are shown in Table 3.

[0121] Table 3 Model fault diagnosis accuracy

[0122] Number of experiments Improving CNNs Traditional CNN 1 100 88.63 2 99.81 88.81 3 99.87 89.19 4 99.69 89.37 5 99.75 89.06 Average accuracy ± standard deviation 99.82±0.11 89.01±0.29

[0123] Table 3 fully demonstrates that, compared to traditional CNNs using the same network architecture, the improved CNN proposed in this invention can more effectively perform fault diagnosis based on motor stator current signals. The average diagnostic accuracy of the improved CNN reached 99.82%, while the average diagnostic accuracy of the traditional CNN was only 89.01%. Furthermore, across five tests, the improved CNN had a smaller standard deviation of accuracy, demonstrating its greater generalization capability. It can be concluded that, when diagnosing gearboxes using motor stator current signals, the improved CNN proposed in this invention can significantly improve the accuracy and stability of gearbox fault identification compared to traditional CNNs, effectively completing end-to-end fault diagnosis tasks.

[0124] Furthermore, using the test set of the third experiment, we used the auxiliary visualization method Figure 7 and Figure 8 Analyze the fault diagnosis results. Figure 7 Is the confusion matrix diagram, from Figure 7 It can be found that when there are 100 samples of each type, only 2 samples of category 16 are misclassified into category 4, and samples of other categories are correctly classified; Figure 8This is a t-SNE dimensionality reduction visualization diagram. It shows that the boundaries of the clusters for category 4 and category 16 are confused, indicating that a small number of samples in these two categories are not accurately classified. The clustering effect of other categories is excellent, with no inter-cluster confusion, which is consistent with the analysis results of the confusion matrix diagram. The analysis results show that the present invention can effectively cluster fault samples of the same category and can also effectively classify fault samples of different categories.

[0125] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. An intelligent gearbox fault diagnosis method based on motor stator current signals, characterized by: The steps include: Step 1: Build a CNN model sample library, which includes the following steps: Step 1-1. Constructing gearbox fault types: The gearbox of a rotating machine includes an input gear and an output gear that mesh with each other. Design a number of input gears with different fault types and b number of output gears with different fault types. By cross-combining a number of input gears and b number of output gears, a*b number of gearbox fault types with different fault types are obtained. Where a and b are both natural numbers. Step 1-2, collecting the motor stator current: The gearbox for each fault type constructed in step 1-1 is sequentially installed in a rotating machine containing a motor; the rotating machine is operated, and the stator current of the motor is collected at a set frequency using a current sensor; the motor stator current signal of the gearbox for each fault type is obtained; Step 1-3, constructing a CNN model sample library: the stator current signal of the gearbox motor for each fault type obtained in step 1-2 constitutes the input sample library of the CNN model; the a*b types of gearbox fault types obtained in step 1-1 form the output sample library of the CNN model; the input sample library and the output sample library together form the CNN model sample library; Step 2: Construct a comb filter, which specifically includes the following steps: Step 2-1: Characterize the fault characteristic frequency f in the current signal fault , the specific expression is: f fault =|f s ±kf fp | When the gearbox is a single-stage gearbox, where: f fp =f r，in or f fp =f r，out Where, f s The frequency of power supply to the motor; f fp The rotation frequency of the faulty gear; k is the harmonic order, a natural number; f r，in is the input gear rotation frequency; f r，out is the output gear rotation frequency; Step 2-2, characterize the noise frequency: the noise frequency includes f s and f s ±c; where c is the torque fluctuation frequency caused by non-gearbox faults; Step 2-3, construct a comb filter: The passband of the comb filter can make the fault characteristic frequency f characterized in step 2-1 fault By,the stop band of the comb filter can suppress the noise frequency characterized by step 2-2; Step 3. Build a CNN model: The CNN model consists of an input layer, a hidden layer, and an output layer. The input layer is the motor stator current signal corresponding to the gearbox. The output layer is the gearbox fault type. The hidden layer includes several convolutional layers. The first convolutional layer has several convolution kernels. Each convolution kernel in the first convolutional layer is the comb filter constructed in step 2. Step 4: Train the CNN model: Use the CNN model sample library constructed in step 1 to train the CNN model sample library built in step 3 to obtain a CNN model with fault diagnosis capabilities; Step 5: Gearbox fault diagnosis, specifically including the following steps: Step 5-1, obtaining the actual motor stator current signal: using a current sensor to collect the motor stator current corresponding to the gearbox to be tested at a set frequency; obtaining the actual motor stator current signal of the gearbox to be tested; Step 5-2, gearbox fault diagnosis: The actual motor stator current signal obtained in step 5-1 is used as input and substituted into the CNN model with fault diagnosis capability obtained in step 4. The output is the fault type of the gearbox to be detected.

2. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 1 is characterized in that: In step 1-1, a=b=4, and the four different fault types of the input gear are severe pitting, moderate pitting, mild pitting, and no fault. Severe pitting, moderate pitting, and mild pitting mean that all the input gears have pitting, and the number of pitting decreases gradually. The four different fault types of the output gear are severe wear, moderate wear, mild wear and no fault; among them, severe wear, moderate wear and mild wear mean that at least one tooth of the output gear is worn, and the wear thickness gradually decreases.

3. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 1 is characterized in that: In step 2-3, the comb filter is constructed using the Fourier kernel function; the time domain function of the comb filter is about time t, f h，m and f l，m function; f h，m is the high cutoff frequency of the mth comb tooth, f l，m is the lower cutoff frequency of the mth comb tooth, 1≤m≤M; M is the number of comb teeth of the comb filter; the stopband bandwidth between adjacent comb teeth is marked as f d .

4. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 3 is characterized in that: f h，m 、f l，m and f d All of them are self-learning parameters. By adjusting the high cutoff frequency f of the first comb tooth h，1 , the low cutoff frequency f of the first comb tooth l，1 and f d Initial values are set for both, and self-learning is performed during the CNN model training in step 4.

5. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 4 is characterized in that: f l，1 The initial value range is [0, 80% × f s / 2];f d The initial value f d0 The frequency resolution of not less than 2 comb filters; f h，1 The initial value range is [f d0 , 80%×f s / 2+f d0 ].

6. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 1, characterized in that: In step 3, the number of convolution kernels in the first convolution layer does not exceed 45, and the average accuracy of gearbox fault diagnosis in step 5 is not less than 99%.

7. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 6, characterized in that: The number of convolution kernels in the first convolution layer is 40.

8. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 1, characterized in that: Before the gearbox fault diagnosis in step 5, first set the motor power supply frequency f corresponding to the gearbox to be tested. s The motor power supply frequency f in step 2 s Compare them. If the two are the same, proceed to step 5. Otherwise, according to the motor power supply frequency f corresponding to the gearbox to be tested, s , repeat steps 1 to 4 to train a new CNN model; then proceed to step 5 based on the new CNN model.

9. The gearbox fault intelligent diagnosis method based on motor stator current signal according to claim 1, characterized in that: In step 2-3, the comb filter is truncated using a Hamming window.

Images